Move function requiring AffinePoint from EdwardsPoint to AffinePoint

[daxpedda]

A bunch of functionality implemented on EdwardsPoint internally just converts to AffinePoint. This functionality should rather live in AffinePoint.

This PR also implement BatchNormalize for EdwardsPoint, which wouldn't have much of a point without these changes to begin with.

[tarcieri]

This is definitely pushing my knowledge of both Edwards curves and Ed448-Goldilocks, and sent me on a rabbit hole reading the paper.

I think both versions of edwards_isogeny are naively implemented, and we would ideally be converting directly from EdwardsPoint to TwistedExtendedPoint (sidebar: which should probably be called TwistedEdwardsPoint for consistency, I guess the former was renamed from ExtendedPoint at some point) without first converting to an intermediate AffinePoint with its associated inversions. There's even a comment to that effect:

So what we can do is optimise each respective isogeny method for a=1 or a = -1 (currently, I just made it really slow and simple)

Edit: opened an issue about this #1349, and it seems to be the reason you weren't seeing speedups using twisted scalar multiplication in #1337

[tarcieri]

To expound on that:

• EdwardsPoint should probably remain the primary user-facing type with the majority of the functionality
• AffinePoint and TwistedExtendedPoint are implementation details

Conversions look like:

CompressedEdwardsY <-> AffinePoint <-> EdwardsPoint <-> TwistedExtendedPoint

...but I think it's fine to preserve methods like compress and decompress directly on EdwardsPoint, but with the actual point compression/decompression implementation factored onto AffinePoint, and the methods on EdwardsPoint delegating to AffinePoint and doing affine conversions as needed.

Likewise, scalar multiplication would ideally do a more direct conversion from EdwardsPoint to TwistedExtendedPoint, and delegate to the implementation there.

[daxpedda]

compress() and decompress() remain available on EdwardsPoint through GroupEncoding. Eventually we should remove these functions entirely and implement GroupEncoding on AffinePoint as well (tracking that in #1348).

I hope I'm on the right track here? I'm kinda trying to align things with the other RustCrypto crates, where most types have very few methods and its all mostly implemented through traits. It also makes sense to me, because otherwise we have all those duplicate methods.

[daxpedda]

Putting on draft until I rebase. Still working on those optimized implementations from #1316.